Active Contributors: Alberto Rota, Carlo Sartori, Federico Monterosso, Iva Milojkovic
Supervisors: Prof. Pietro Cerveri, PhD; Matteo Rossi, PhD
A Deep Learning model for superresolution of CT Images. This superresolution task consists in cthe combination of denoising and increasing of spatial extent. Both a higher PSNR and SSIM were archieved when compared to standard non-ML superresolution strategies like bilinear, cubic and quintic interpolation: such methods avereaged a PSNR of 35.2dB and SSIM of 0.88 on the available dataset. The developed dense model reached a 43dB PSNR and 0.95 SSIM on the test set (20% of the total dataset size) only
Input: 128x128x64 low-res CT-scans, with superimposed gaussian noise
Output: 256x256x64 high-res CT-scans, denoised
The network implements a variation on the standard DenseNet architecture, where a 4x upsampling followed by a convolutional layer with stride 2 is added at the beginning. The model has ~45K parameters and trains 1 epoch in 102 seconds on an NVIDIA P100 GPU. An efficient training earlystopped after 55 of the 100 desired epoch, while using a batch size of 4 slices to avoid overloading the RAM. Even tho multiple ad hoc loss functions were tested (which took in consideration the PSNR), the best performance was obtained from MSE.
The model output can be viewed sliced or in 3D from the provided MATLAB script